Merge branch 'screwaround'

Makefile

@@ -9,7 +9,7 @@ CYGWIN  = $(findstring CYGWIN,  $(UNAME))
 MINGW   = $(findstring MINGW32, $(UNAME))
 
 CC     ?= clang
-CFLAGS += -Wall -Wextra -I. -fno-strict-aliasing -fsigned-char
+CFLAGS += -Wall -Wextra -I. -fno-strict-aliasing -fsigned-char -O2
 CFLAGS += -DGMQCC_GITINFO="`git describe`"
 #turn on tons of warnings if clang is present
 # but also turn off the STUPID ONES
@@ -101,14 +101,15 @@ check: all
 clean:
 	rm -f *.o $(GMQCC) $(QCVM) $(TESTSUITE) *.dat
 
-# deps
+depend:
-$(OBJ_D) $(OBJ_C) $(OBJ_X): gmqcc.h opts.def
+	makedepend -Y -w 65536 \
-main.o:   lexer.h
+		$(subst .o,.c,$(OBJ_D))
-parser.o: ast.h lexer.h
+	makedepend -a -Y -w 65536 \
-ftepp.o:  lexer.h
+		$(subst .o,.c,$(OBJ_T))
-lexer.o:  lexer.h
+	makedepend -a -Y -w 65536 \
-ast.o:    ast.h ir.h
+		$(subst .o,.c,$(OBJ_C))
-ir.o:     ir.h
+	makedepend -a -Y -w 65536 \
+		$(subst .o,.c,$(OBJ_X))
 
 #install rules
 install: install-gmqcc install-qcvm install-doc
@@ -122,3 +123,30 @@ install-doc:
 	install -d -m755               $(DESTDIR)$(MANDIR)/man1
 	install    -m755  doc/gmqcc.1  $(DESTDIR)$(MANDIR)/man1/
 	install    -m755  doc/qcvm.1   $(DESTDIR)$(MANDIR)/man1/
+
+# DO NOT DELETE
+
+util.o: gmqcc.h opts.def
+code.o: gmqcc.h opts.def
+ast.o: gmqcc.h opts.def ast.h ir.h
+ir.o: gmqcc.h opts.def ir.h
+conout.o: gmqcc.h opts.def
+ftepp.o: gmqcc.h opts.def lexer.h
+opts.o: gmqcc.h opts.def
+file.o: gmqcc.h opts.def
+utf8.o: gmqcc.h opts.def
+correct.o: gmqcc.h opts.def
+
+test.o: gmqcc.h opts.def
+util.o: gmqcc.h opts.def
+conout.o: gmqcc.h opts.def
+file.o: gmqcc.h opts.def
+
+main.o: gmqcc.h opts.def lexer.h
+lexer.o: gmqcc.h opts.def lexer.h
+parser.o: gmqcc.h opts.def lexer.h ast.h ir.h
+file.o: gmqcc.h opts.def
+
+util.o: gmqcc.h opts.def
+conout.o: gmqcc.h opts.def
+file.o: gmqcc.h opts.def

correct.c

@@ -1,6 +1,7 @@
 /*
  * Copyright (C) 2012, 2013
  *     Dale Weiler
+ *     Wolfgang Bumiller
  * 
  * Permission is hereby granted, free of charge, to any person obtaining a copy of
  * this software and associated documentation files (the "Software"), to deal in
@@ -38,13 +39,25 @@
  *  There is actually no way to know for sure that certian identifers
  *  such as "lates", need to be corrected to "late" or "latest" or any
  *  other permutations that look lexically the same.  This is why we
- *  must advocate the usage of probabilities.  This implies that we're
+ *  must advocate the usage of probabilities.  This means that instead of
- *  trying to find the correction for C, out of all possible corrections
+ *  just guessing, instead we're trying to find the correction for C,
- *  that maximizes the probability of C for the original identifer I.
+ *  out of all possible corrections that maximizes the probability of C
+ *  for the original identifer I.
  *
- *  Bayes' Therom suggests something of the following:
+ *  Thankfully there exists some theroies for probalistic interpretations
+ *  of data.  Since we're operating on two distictive intepretations, the
+ *  transposition from I to C. We need something that can express how much
+ *  degree of I should rationally change to become C.  this is called the
+ *  Bayesian interpretation. You can read more about it from here:
+ *  http://www.celiagreen.com/charlesmccreery/statistics/bayestutorial.pdf
+ *  (which is probably the only good online documentation for bayes theroy
+ *  no lie.  Everything else just sucks ..)  
+ * 
+ *  Bayes' Thereom suggests something like the following:
  *      AC P(I|C) P(C) / P(I)
- *  Since P(I) is the same for every possibly I, we can ignore it giving
+ * 
+ *  However since P(I) is the same for every possibility of I, we can
+ *  completley ignore it giving just:
  *      AC P(I|C) P(C)
  *
  *  This greatly helps visualize how the parts of the expression are performed
@@ -56,12 +69,9 @@
  *  2: P(I|C), the probability that I would be used, when the programmer
  *     really meant C.  This is the error model.
  *
- *  3: AC, the control mechanisim, which implies the enumeration of all
+ *  3: AC, the control mechanisim, an enumerator if you will, one that
- *     feasible values of C, and then determine the one that gives the
+ *     enumerates all feasible values of C, to determine the one that
- *     greatest probability score. Selecting it as the "correction"
+ *     gives the greatest probability score.
- *   
- *
- * The requirement for complex expression involving two models:
  * 
  *  In reality the requirement for a more complex expression involving
  *  two seperate models is considerably a waste.  But one must recognize
@@ -70,29 +80,230 @@
  *  estimate P(C|I) you have to consider both the probability of C and
  *  probability of the transposition from C to I.  It's simply much more
  *  cleaner, and direct to seperate the two factors.
+ *
+ *  Research tells us that 80% to 95% of all spelling errors have an edit
+ *  distance no greater than one.  Knowing this we can optimize for most
+ *  cases of mistakes without taking a performance hit.  Which is what we
+ *  base longer edit distances off of.  Opposed to the original method of
+ *  I had concieved of checking everything.     
+ *  
+ * A little information on additional algorithms used:
+ *
+ *   Initially when I implemented this corrector, it was very slow.
+ *   Need I remind you this is essentially a brute force attack on strings,
+ *   and since every transformation requires dynamic memory allocations,
+ *   you can easily imagine where most of the runtime conflated.  Yes
+ *   It went right to malloc.  More than THREE MILLION malloc calls are
+ *   performed for an identifier about 16 bytes long.  This was such a
+ *   shock to me.  A forward allocator (or as some call it a bump-point
+ *   allocator, or just a memory pool) was implemented. To combat this.
+ *
+ *   But of course even other factors were making it slow.  Initially
+ *   this used a hashtable.  And hashtables have a good constant lookup
+ *   time complexity.  But the problem wasn't in the hashtable, it was
+ *   in the hashing (despite having one of the fastest hash functions
+ *   known).  Remember those 3 million mallocs? Well for every malloc
+ *   there is also a hash.  After 3 million hashes .. you start to get
+ *   very slow.  To combat this I had suggested burst tries to Blub.
+ *   The next day he had implemented them. Sure enough this brought
+ *   down the runtime by a factor > 100%
+ *
+ *   The trie initially was designed to work on all strings, but later it
+ *   became aparent that not only was this not a requirement. It was also
+ *   slowing down get/sets' for the trie.  To fully understand, only
+ *   correct_alpha needs to be understood by the trie system, knowing this
+ *   We can combat the slowness using a very clever but evil optimization.
+ *   By Setting a fixed sized amount of branches for the trie using a
+ *   char-to-index map into the branches. We've complelty made the trie
+ *   accesses entierly constant in lookup time.  No really, a lookup is
+ *   literally trie[str[0]] [str[1]] [2] .... .value.
+ * 
+ * 
+ * Future Work (If we really need it)
+ *
+ *   Currently we can only distinguish one source of error in the
+ *   language model we use.  This could become an issue for identifiers
+ *   that have close colliding rates, e.g colate->coat yields collate.
+ *
+ *   Currently the error model has been fairly trivial, the smaller the
+ *   edit distance the smaller the error.  This usually causes some un-
+ *   expected problems. e.g reciet->recite yields recipt.  For QuakeC
+ *   this could become a problem when lots of identifiers are involved. 
+ *
+ *   Our control mechanisim could use a limit, i.e limit the number of
+ *   sets of edits for distance X.  This would also increase execution
+ *   speed considerably.
  */
 
-/* some hashtable management for dictonaries */
+
-static size_t *correct_find(ht table, const char *word) {
+#define CORRECT_POOL_SIZE (128*1024*1024)
-    return (size_t*)util_htget(table, word);
+/*
+ * A forward allcator for the corrector.  This corrector requires a lot
+ * of allocations.  This forward allocator combats all those allocations
+ * and speeds us up a little.  It also saves us space in a way since each
+ * allocation isn't wasting a little header space for when NOTRACK isn't
+ * defined.
+ */
+static unsigned char **correct_pool_data = NULL;
+static unsigned char  *correct_pool_this = NULL;
+static size_t          correct_pool_addr = 0;
+
+static GMQCC_INLINE void correct_pool_new(void) {
+    correct_pool_addr = 0;
+    correct_pool_this = (unsigned char *)mem_a(CORRECT_POOL_SIZE);
+
+    vec_push(correct_pool_data, correct_pool_this);
 }
 
-static int correct_update(ht *table, const char *word) {
+static GMQCC_INLINE void *correct_pool_alloc(size_t bytes) {
-    size_t *data = correct_find(*table, word);
+    void *data;
-    if (!data)
+    if (correct_pool_addr + bytes>= CORRECT_POOL_SIZE)
-        return 0;
+        correct_pool_new();
 
-    (*data)++;
+    data               = (void*)correct_pool_this;
-    return 1;
+    correct_pool_this += bytes;
+    correct_pool_addr += bytes;
+    return data;
 }
 
+static GMQCC_INLINE void correct_pool_delete(void) {
+    size_t i;
+    for (i = 0; i < vec_size(correct_pool_data); ++i)
+        mem_d(correct_pool_data[i]);
+
+    correct_pool_data = NULL;
+    correct_pool_this = NULL;
+    correct_pool_addr = 0;
+}
+
+
+static GMQCC_INLINE char *correct_pool_claim(const char *data) {
+    char *claim = util_strdup(data);
+    correct_pool_delete();
+    return claim;
+}
 
 /*
  * _ is valid in identifiers. I've yet to implement numerics however
  * because they're only valid after the first character is of a _, or
  * alpha character.
  */
-static const char correct_alpha[] = "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ_";
+static const char correct_alpha[] = "abcdefghijklmnopqrstuvwxyz"
+                                    "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+                                    "_"; /* TODO: Numbers ... */
+
+static const size_t correct_alpha_index[0x80] = {
+     0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+     0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+     0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+     0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+     0,  0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14,
+    15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25,  0,  0,  0,  0, 52,
+     0, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
+    41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51,  0,  0,  0,  0,  0
+};
+
+/*
+ * A fast space efficent trie for a dictionary of identifiers.  This is
+ * faster than a hashtable for one reason.  A hashtable itself may have
+ * fast constant lookup time, but the hash itself must be very fast. We
+ * have one of the fastest hash functions for strings, but if you do a
+ * lost of hashing (which we do, almost 3 million hashes per identifier)
+ * a hashtable becomes slow.
+ */
+correct_trie_t* correct_trie_new() {
+    correct_trie_t *t = (correct_trie_t*)mem_a(sizeof(correct_trie_t));
+    t->value   = NULL;
+    t->entries = NULL;
+    return t;
+}
+
+void correct_trie_del_sub(correct_trie_t *t) {
+    size_t i;
+    if (!t->entries)
+        return;
+    for (i = 0; i < sizeof(correct_alpha)-1; ++i) {
+        correct_trie_del_sub(&t->entries[i]);
+    }
+    mem_d(t->entries);
+}
+
+void correct_trie_del(correct_trie_t *t) {
+    size_t i;
+    if (t->entries) {
+        for (i = 0; i < sizeof(correct_alpha)-1; ++i)
+            correct_trie_del_sub(&t->entries[i]);
+        mem_d(t->entries);
+    }
+    mem_d(t);
+}
+
+void* correct_trie_get(const correct_trie_t *t, const char *key) {
+    const unsigned char *data = (const unsigned char*)key;
+
+    while (*data) {
+        if (!t->entries)
+            return NULL;
+        t = t->entries + correct_alpha_index[*data];
+        ++data;
+    }
+    return t->value;
+}
+
+void correct_trie_set(correct_trie_t *t, const char *key, void * const value) {
+    const unsigned char *data = (const unsigned char*)key;
+    while (*data) {
+        if (!t->entries) {
+            t->entries = (correct_trie_t*)mem_a(sizeof(correct_trie_t)*(sizeof(correct_alpha)-1));
+            memset(t->entries, 0, sizeof(correct_trie_t)*(sizeof(correct_alpha)-1));
+        }
+        t = t->entries + correct_alpha_index[*data];
+        ++data;
+    }
+    t->value = value;
+}
+
+
+/*
+ * Implementation of the corrector algorithm commences. A very efficent
+ * brute-force attack (thanks to tries and mempool :-)).
+ */
+static GMQCC_INLINE size_t *correct_find(correct_trie_t *table, const char *word) {
+    return (size_t*)correct_trie_get(table, word);
+}
+
+static GMQCC_INLINE bool correct_update(correct_trie_t* *table, const char *word) {
+    size_t *data = correct_find(*table, word);
+    if (!data)
+        return false;
+
+    (*data)++;
+    return true;
+}
+
+void correct_add(correct_trie_t* table, size_t ***size, const char *ident) {
+    size_t     *data = NULL;
+    const char *add  = ident;
+
+    if (!correct_update(&table, add)) {
+        data  = (size_t*)mem_a(sizeof(size_t));
+        *data = 1;
+
+        vec_push((*size), data);
+        correct_trie_set(table, add, data);
+    }
+}
+
+void correct_del(correct_trie_t* dictonary, size_t **data) {
+    size_t       i;
+    const size_t vs = vec_size(data);
+
+    for (i = 0; i < vs; i++)
+        mem_d(data[i]);
+
+    vec_free(data);
+    correct_trie_del(dictonary);
+}
 
 /*
  * correcting logic for the following forms of transformations:
@@ -100,73 +311,81 @@ static const char correct_alpha[] = "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQ
  *  2) transposition
  *  3) alteration
  *  4) insertion
+ *
+ * These functions could take an additional size_t **size paramater
+ * and store back the results of their new length in an array that
+ * is the same as **array for the memcmp in correct_exists. I'm just
+ * not able to figure out how to do that just yet.  As my brain is
+ * not in the mood to figure out that logic.  This is a reminder to
+ * do it, or for someone else to :-) correct_edit however would also
+ * need to take a size_t ** to carry it along (would all the argument
+ * overhead be worth it?)  
  */
-static size_t correct_deletion(const char *ident, char **array, size_t index) {
+static size_t correct_deletion(const char *ident, char **array) {
-    size_t itr;
+    size_t       itr = 0;
-    size_t len = strlen(ident);
+    const size_t len = strlen(ident);
 
-    for (itr = 0; itr < len; itr++) {
+    for (; itr < len; itr++) {
-        char *a = (char*)mem_a(len+1);
+        char *a = (char*)correct_pool_alloc(len+1);
         memcpy(a, ident, itr);
         memcpy(a + itr, ident + itr + 1, len - itr);
-        array[index + itr] = a;
+        array[itr] = a;
     }
 
     return itr;
 }
 
-static size_t correct_transposition(const char *ident, char **array, size_t index) {
+static size_t correct_transposition(const char *ident, char **array) {
-    size_t itr;
+    size_t       itr = 0;
-    size_t len = strlen(ident);
+    const size_t len = strlen(ident);
 
-    for (itr = 0; itr < len - 1; itr++) {
+    for (; itr < len - 1; itr++) {
         char  tmp;
-        char *a = (char*)mem_a(len+1);
+        char *a = (char*)correct_pool_alloc(len+1);
         memcpy(a, ident, len+1);
         tmp      = a[itr];
         a[itr  ] = a[itr+1];
         a[itr+1] = tmp;
-        array[index + itr] = a;
+        array[itr] = a;
     }
 
     return itr;
 }
 
-static size_t correct_alteration(const char *ident, char **array, size_t index) {
+static size_t correct_alteration(const char *ident, char **array) {
-    size_t itr;
+    size_t       itr = 0;
-    size_t jtr;
+    size_t       jtr = 0;
-    size_t ktr;
+    size_t       ktr = 0;
-    size_t len    = strlen(ident);
+    const size_t len = strlen(ident);
 
-    for (itr = 0, ktr = 0; itr < len; itr++) {
+    for (; itr < len; itr++) {
         for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++, ktr++) {
-            char *a = (char*)mem_a(len+1);
+            char *a = (char*)correct_pool_alloc(len+1);
             memcpy(a, ident, len+1);
             a[itr] = correct_alpha[jtr];
-            array[index + ktr] = a;
+            array[ktr] = a;
         }
     }
 
     return ktr;
 }
 
-static size_t correct_insertion(const char *ident, char **array, size_t index) {
+static size_t correct_insertion(const char *ident, char **array) {
-    size_t itr;
+    size_t       itr = 0;
-    size_t jtr;
+    size_t       jtr = 0;
-    size_t ktr;
+    const size_t len = strlen(ident);
-    size_t len    = strlen(ident);
 
-    for (itr = 0, ktr = 0; itr <= len; itr++) {
+    for (; itr <= len; itr++) {
-        for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++, ktr++) {
+        for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++) {
-            char *a = (char*)mem_a(len+2);
+            char *a = (char*)correct_pool_alloc(len+2);
             memcpy(a, ident, itr);
-            a[itr] = correct_alpha[jtr];
             memcpy(a + itr + 1, ident + itr, len - itr + 1);
-            array[index + ktr] = a;
+            a[itr] = correct_alpha[jtr];
+            array[itr * (sizeof(correct_alpha)-1) + jtr] = a;
         }
     }
 
-    return ktr;
+    return (len+1)*(sizeof(correct_alpha)-1);
 }
 
 static GMQCC_INLINE size_t correct_size(const char *ident) {
@@ -183,15 +402,15 @@ static GMQCC_INLINE size_t correct_size(const char *ident) {
 
 static char **correct_edit(const char *ident) {
     size_t next;
-    char **find = (char**)mem_a(correct_size(ident) * sizeof(char*));
+    char **find = (char**)correct_pool_alloc(correct_size(ident) * sizeof(char*));
 
     if (!find)
         return NULL;
 
-    next  = correct_deletion     (ident, find, 0);
+    next  = correct_deletion     (ident, find);
-    next += correct_transposition(ident, find, next);
+    next += correct_transposition(ident, find+next);
-    next += correct_alteration   (ident, find, next);
+    next += correct_alteration   (ident, find+next);
-    /*****/ correct_insertion    (ident, find, next);
+    /*****/ correct_insertion    (ident, find+next);
 
     return find;
 }
@@ -203,48 +422,86 @@ static char **correct_edit(const char *ident) {
  */
 static int correct_exist(char **array, size_t rows, char *ident) {
     size_t itr;
-    for (itr = 0; itr < rows; itr++)
+    /*
-        if (!strcmp(array[itr], ident))
+     * As an experiment I tried the following assembly for memcmp here:
+     *
+     * correct_cmp_loop: 
+     * incl %eax            ; eax =  LHS
+     * incl %edx            ; edx =  LRS
+     * cmpl %eax, %ebx      ; ebx = &LHS[END_POS]
+     *
+     * jbe correct_cmp_eq
+     * movb (%edx), %cl     ; micro-optimized even on atoms :-)
+     * cmpb %cl, (%eax)     ; ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+     * jg  correct_cmp_gt
+     * jge correct_cmp_loop
+     * ...
+     *
+     * Despite how much optimization went in to this, the speed was
+     * being conflicted by the strlen(ident) used for &LHS[END_POS]
+     * If we could eliminate the strlen with what I suggested on line
+     * 311 ... we can accelerate this whole damn thing quite a bit.
+     *
+     * However there is still something we can do here that does give
+     * us a little more speed.  Although one more branch, we know for
+     * sure there is at least one byte to compare, if that one byte
+     * simply isn't the same we can skip the full check. Which means
+     * we skip a whole strlen call.
+     */
+    for (itr = 0; itr < rows; itr++) {
+        if (!memcmp(array[itr], ident, strlen(ident)))
             return 1;
+    }
 
     return 0;
 }
 
-static char **correct_known(ht table, char **array, size_t rows, size_t *next) {
+static GMQCC_INLINE char **correct_known_resize(char **res, size_t *allocated, size_t size) {
-    size_t itr;
+    size_t oldallocated = *allocated;
-    size_t jtr;
+    char **out;
-    size_t len;
+    if (size < oldallocated)
-    size_t row;
+        return res;
-    char **res = NULL;
-    char **end;
 
-    for (itr = 0, len = 0; itr < rows; itr++) {
+    out = correct_pool_alloc(sizeof(*res) * oldallocated + 32);
+    memcpy(out, res, sizeof(*res) * oldallocated);
+
+    *allocated += 32;
+    return out;
+}
+
+static char **correct_known(correct_trie_t* table, char **array, size_t rows, size_t *next) {
+    size_t itr = 0;
+    size_t jtr = 0;
+    size_t len = 0;
+    size_t row = 0;
+    size_t nxt = 8;
+    char **res = correct_pool_alloc(sizeof(char *) * nxt);
+    char **end = NULL;
+
+    for (; itr < rows; itr++) {
         end = correct_edit(array[itr]);
         row = correct_size(array[itr]);
 
+        /* removing jtr=0 here speeds it up by 100ms O_o */
         for (jtr = 0; jtr < row; jtr++) {
             if (correct_find(table, end[jtr]) && !correct_exist(res, len, end[jtr])) {
-                res        = mem_r(res, sizeof(char*) * (len + 1));
+                res        = correct_known_resize(res, &nxt, len+1);
                 res[len++] = end[jtr];
-            } else {
-                mem_d(end[jtr]);
             }
         }
-
-        mem_d(end);
     }
 
     *next = len;
     return res;
 }
 
-static char *correct_maximum(ht table, char **array, size_t rows) {
+static char *correct_maximum(correct_trie_t* table, char **array, size_t rows) {
     char   *str = NULL;
     size_t *itm = NULL;
-    size_t  itr;
+    size_t  itr = 0;
-    size_t  top;
+    size_t  top = 0;
 
-    for (itr = 0, top = 0; itr < rows; itr++) {
+    for (; itr < rows; itr++) {
         if ((itm = correct_find(table, array[itr])) && (*itm > top)) {
             top = *itm;
             str = array[itr];
@@ -254,77 +511,37 @@ static char *correct_maximum(ht table, char **array, size_t rows) {
     return str;
 }
 
-static void correct_cleanup(char **array, size_t rows) {
-    size_t itr;
-    for (itr = 0; itr < rows; itr++)
-        mem_d(array[itr]);
-
-    mem_d(array);
-}
-
 /*
  * This is the exposed interface:
  * takes a table for the dictonary a vector of sizes (used for internal
- * probability calculation, and an identifier to "correct"
+ * probability calculation), and an identifier to "correct".
- *
- * the add function works the same.  Except the identifier is used to
- * add to the dictonary.  
  */
-void correct_add(ht table, size_t ***size, const char *ident) {
+char *correct_str(correct_trie_t* table, const char *ident) {
-    size_t     *data = NULL;
+    char **e1      = NULL;
-    const char *add  = ident;
+    char **e2      = NULL;
-    
+    char  *e1ident = NULL;
-    if (!correct_update(&table, add)) {
+    char  *e2ident = NULL;
-        data  = (size_t*)mem_a(sizeof(size_t));
-        *data = 1;
-
-        vec_push((*size), data);
-        util_htset(table, add, data);
-    }
-}
-
-char *correct_str(ht table, const char *ident) {
-    char **e1;
-    char **e2;
-    char  *e1ident;
-    char  *e2ident;
-    char  *found = util_strdup(ident);
-
     size_t e1rows  = 0;
     size_t e2rows  = 0;
 
+    correct_pool_new();
+
     /* needs to be allocated for free later */
     if (correct_find(table, ident))
-        return found;
+        return correct_pool_claim(ident);
 
     if ((e1rows = correct_size(ident))) {
         e1      = correct_edit(ident);
 
-        if ((e1ident = correct_maximum(table, e1, e1rows))) {
+        if ((e1ident = correct_maximum(table, e1, e1rows)))
-            mem_d(found);
+            return correct_pool_claim(e1ident);
-            found = util_strdup(e1ident);
-            correct_cleanup(e1, e1rows);
-            return found;
-        }
     }
 
     e2 = correct_known(table, e1, e1rows, &e2rows);
-    if (e2rows && ((e2ident = correct_maximum(table, e2, e2rows)))) {
+    if (e2rows && ((e2ident = correct_maximum(table, e2, e2rows))))
-        mem_d(found);
+        return correct_pool_claim(e2ident);
-        found = util_strdup(e2ident);
-    }
 
-    correct_cleanup(e1, e1rows);
-    correct_cleanup(e2, e2rows);
 
-    return found;
+    correct_pool_delete();
-}
+    return util_strdup(ident);
-
-void correct_del(ht dictonary, size_t **data) {
-    size_t i;
-    for (i = 0; i < vec_size(data); i++)
-        mem_d(data[i]);
-
-    vec_free(data);
-    util_htdel(dictonary);
 }

gmqcc.h

@@ -315,6 +315,13 @@ void _util_vec_grow(void **a, size_t i, size_t s);
 #define vec_upload(X,Y,S) memcpy(vec_add((X), (S) * sizeof(*(Y))), (Y), (S) * sizeof(*(Y)))
 #define vec_remove(A,I,N) memmove((A)+(I),(A)+((I)+(N)),sizeof(*(A))*(vec_meta(A)->used-(I)-(N))),vec_meta(A)->used-=(N)
 
+typedef struct trie_s {
+    void          *value;
+    struct trie_s *entries;
+} correct_trie_t;
+
+correct_trie_t* correct_trie_new();
+
 typedef struct hash_table_t {
     size_t                size;
     struct hash_node_t **table;
@@ -426,9 +433,9 @@ GMQCC_INLINE FILE   *file_open   (const char *, const char *);
 /*===================================================================*/
 /*=========================== correct.c =============================*/
 /*===================================================================*/
-void  correct_del(ht, size_t **);
+void  correct_del(correct_trie_t*, size_t **);
-void  correct_add(ht, size_t ***, const char *);
+void  correct_add(correct_trie_t*, size_t ***, const char *);
-char *correct_str(ht, /********/  const char *);
+char *correct_str(correct_trie_t*, /********/  const char *);
 
 /*===================================================================*/
 /*=========================== code.c ================================*/

parser.c

@@ -75,7 +75,7 @@ typedef struct {
     ht *typedefs;
 
     /* same as above but for the spelling corrector */
-    ht       *correct_variables;
+    correct_trie_t  **correct_variables;
     size_t         ***correct_variables_score;  /* vector of vector of size_t* */
 
     /* not to be used directly, we use the hash table */
@@ -1999,7 +1999,7 @@ static void parser_enterblock(parser_t *parser)
     vec_push(parser->_block_ctx, parser_ctx(parser));
 
     /* corrector */
-    vec_push(parser->correct_variables, util_htnew(PARSER_HT_SIZE));
+    vec_push(parser->correct_variables, correct_trie_new());
     vec_push(parser->correct_variables_score, NULL);
 }